JP2010078650A - Speech recognizer and method thereof - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognizer stably recognizing speech even under noise. <P>SOLUTION: The speech recognizer executes: extracting noisy speech feature vectors from input noisy speech in each frame; estimating noise feature distribution parameters of noise feature vectors concerning noise superposed on the noisy speech; calculating combination Gaussian distribution parameters of clean speech feature vectors and the noisy speech feature vectors by using unscented transform from noise feature distribution parameters and antedating distribution parameters of previously stored clean speech feature vectors; calculating ex post facto distribution parameters of the clean speech feature vectors from the noisy speech feature vectors by using the combination Gaussian distribution parameters; comparing ex post facto distribution parameters with previously stored word standard patterns in each frame; and outputting word strings of noisy speech based on the comparison result. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a speech recognition apparatus and method for recognizing speech uttered under noise.

  Declining speech recognition performance under noise is one of the major problems with speech recognition systems. There is a “speech enhancement method” as a method for improving resistance to noise (noise) of a speech recognition system. This speech enhancement method is a method for estimating clean speech from noisy speech in which noise is superimposed on clean speech. In particular, a method for estimating clean speech in a speech feature area is referred to as “speech feature enhancement method” or “feature enhancement method”.

  A speech recognition apparatus capable of realizing this feature enhancement method operates as follows.

  First, the speech recognition apparatus extracts a noisy speech feature vector from noisy speech with noise superimposed.

  Next, the speech recognition apparatus estimates a clean speech feature vector from a noisy speech feature vector.

  Finally, the speech recognition apparatus collates the estimated clean speech feature vector with a standard pattern of words, and outputs a recognition result word string.

  Non-Patent Document 1 discloses a feature enhancement method that applies the properties of a coupled Gaussian distribution. In this feature enhancement method, it is assumed that a clean speech feature vector and a noisy speech feature vector have a joint Gaussian distribution, and parameters of this joint Gaussian distribution are known. This feature enhancement method calculates the posterior mean and posterior covariance of the clean speech feature vector when the noisy speech feature vector is observed.

  Here, the problem is how to calculate the parameters of this coupled Gaussian distribution. Since the degradation process of the speech feature vector due to noise involves nonlinearity, the estimation of the combined Gaussian distribution parameter becomes a nonlinear estimation problem and cannot be solved analytically.

In the prior art, as disclosed in Non-Patent Document 1, by using first-order Taylor approximation, this nonlinear estimation problem is first replaced with a linear estimation problem, and this linear estimation problem is analyzed, thereby combining Gaussian distributions. Calculate the parameters.
V. Stouten, H. Van hamme, and P. Wambacq, "Model-based feature enhancement with uncertainty decoding for noise robust ASR," Speech Communication, vol. 48, pp. 1502-1514, 2006.

  However, since the above conventional technique linearly approximates a nonlinear function by first-order Taylor expansion, a large approximation error occurs. Therefore, the calculation accuracy of the combined Gaussian distribution parameter is low, and as a result, there is a problem that sufficient speech recognition performance cannot be obtained under noise.

  Accordingly, the present invention provides a speech recognition apparatus and method for solving the above-described problems of the prior art and capable of stably performing speech recognition even under noise.

  The present invention includes a feature extraction unit that extracts a noisy speech feature vector for each frame from an input noisy speech, a noise estimation unit that estimates a noise feature distribution parameter of a noise feature vector related to noise superimposed on the noisy speech, A prior distribution parameter storage unit that stores a prior distribution parameter of a clean speech feature vector related to clean speech, and the noise feature distribution parameter and the prior distribution parameter, using the unscented transform, the clean speech feature vector and the noisy A Gaussian distribution calculation unit that calculates a combined Gaussian distribution parameter of a speech feature vector for each frame, and a posterior distribution parameter of the clean speech feature vector from the noisy speech feature vector using the combined Gaussian distribution parameter. In A calculation execution unit that outputs, a posterior distribution parameter, a pre-stored word standard pattern for each frame, and a collation unit that outputs the noisy speech word string based on the collation result; A speech recognition apparatus comprising:

  According to the present invention, voice recognition can be performed stably even under noise.

  Hereinafter, a speech recognition apparatus 10 according to an embodiment of the present invention will be described with reference to the drawings.

(First embodiment)
A voice recognition device 10 according to a first embodiment will be described with reference to FIGS.

  FIG. 1 is a block diagram of a speech recognition apparatus 10 according to the present embodiment.

  As shown in FIG. 1, the speech recognition apparatus 10 includes a feature extraction unit 11, a noise estimation unit 12, a feature enhancement unit 13, and a collation unit 14.

  The voice recognition device 10 can also be realized by using, for example, a general-purpose computer device as basic hardware. That is, the feature extraction unit 11, the noise estimation unit 12, the feature enhancement unit 13, and the collation unit 14 can be realized by causing a processor mounted on the computer device to execute a program. At this time, the speech recognition apparatus 10 may be realized by installing the above program in a computer device in advance, or may be stored in a storage medium such as a CD-ROM or distributed through the network. Then, this program may be realized by appropriately installing it in a computer device.

  The feature extraction unit 11 will be described.

  The feature extraction unit 11 extracts a vector representing the voice feature from the input noisy voice signal. In this noisy voice, noise is superimposed on the clean voice.

  Specifically, the feature extraction unit 11 receives an audio signal that is noisy speech. Next, the feature extraction unit 11 cuts out a short-time frame (hereinafter simply referred to as a frame) from the audio signal while shifting the cut-out window little by little along the time series. Next, the feature extraction unit 11 converts each frame into a feature vector and outputs a time-series noisy speech feature vector. As the feature vector, for example, an MFCC (Mel-Frequency Cepstral Coefficients) vector is used. Hereinafter, a noisy speech feature vector (hereinafter referred to as a noisy vector) is set as y.

  The noise estimation unit 12 will be described.

  The noise estimation unit 12 estimates a noise feature distribution parameter (hereinafter simply referred to as a noise parameter) of a noise feature vector from the noisy vector y for each frame.

  Specifically, the noise parameter is the mean and covariance of the noise feature vector. For example, the mean and covariance are calculated from a set of feature vectors extracted from a noise-only section (time) that does not include the speech before the start of speech, and thereafter, the noise during speech is assumed to be unchanged, Output this mean and covariance for all frames.

  Also, when it is assumed that the noise fluctuates during utterance, the noise parameter may be updated using the feature vector of the section every time a section not including speech is detected using the voice section detector. Good.

  In the following description, the noise feature vector is set to n. Also, the noise parameters, that is, the mean and covariance of the noise feature vector n are set to μn and Σn, respectively.

  The feature enhancement unit 13 will be described.

  The feature enhancement unit 13 calculates a clean speech feature posterior distribution parameter (hereinafter referred to as a posterior distribution parameter) that is a posterior distribution parameter of the clean speech feature vector (hereinafter referred to as a clean vector) from the noisy vector y and the noise parameter.

  The posterior distribution parameters are specifically the posterior mean and posterior covariance of the clean vector when the noisy vector y is observed.

  In the following description, the clean vector is set to x. Further, the posterior distribution parameters, that is, the posterior mean and posterior covariance of the clean vector x when the noisy vector y is observed are set as μx | y and Σx | y, respectively. Details of the feature enhancement unit 13 will be described later.

  The verification unit 14 will be described.

  The collation unit 14 collates the posterior distribution parameter of the clean vector x with a pre-stored word standard pattern for each frame, and outputs the noisy speech word string based on the collation result.

  Standard Viterbi decoding is performed using the posterior average μx | y calculated by the feature enhancement unit 13 as an estimated value of the clean vector x.

  Non-Patent Document 2 (L. Deng, J. Droppo, and A. Acero, “Dynamic compensation of HMM variances using the feature enhancement uncertainty computed from a parametric model of speech distortion,” IEEE Transactions on Speech and Audio Processing, vol. 13, no. 3, pp. 412-412, May 2005.) Using the posterior mean μx | y and the posterior covariance Σx | y, Also good.

  By performing matching while taking into account the magnitude of the posterior covariance (answerability), frames with a large answer are less reliable and the impact on the matching is reduced. Conversely, frames with a small answer are certain. As a frame, the influence on collation is increased, and the speech recognition performance is improved.

  Next, details of the feature enhancement unit 13 will be described with reference to the block diagram of FIG.

  As shown in FIG. 2, the feature enhancement unit 13 includes a prior distribution parameter storage unit 131, a Gaussian distribution storage unit 132, a Gaussian distribution calculation unit 133, and a calculation execution unit 134.

  The prior distribution parameter storage unit 131 will be described.

  The prior distribution parameter storage unit 131 stores clean speech feature prior distribution parameters of the clean vector x (hereinafter simply referred to as prior distribution parameters).

  Specifically, the prior average μx and the prior covariance Σx of the clean vector x are stored. The prior distribution parameter is calculated in advance using an audio corpus recorded in a quiet environment.

  More specifically, the mean and covariance are calculated from a set of feature vectors extracted from the clean speech corpus. When the speaker or the utterance content is known in advance, the speaker or a corpus specialized for the utterance content can be used.

  Moreover, when a speaker or utterance content is not specified in advance, it is preferable to use various speakers and a corpus including various utterance contents.

  The Gaussian distribution storage unit 132 will be described.

  The Gaussian distribution storage unit 132 stores a combined Gaussian distribution parameter (hereinafter simply referred to as a Gaussian parameter) that is a combined Gaussian distribution parameter of the clean vector x and the noisy vector y. That is, the Gaussian distribution storage unit 132 stores the Gaussian parameters output from the Gaussian distribution calculation unit 133.

  The Gaussian parameters are the prior mean μx and prior covariance Σx of the clean vector x, the mean μy and covariance Σy of the noisy vector y, and the cross covariance Σxy of the clean vector x and the noisy vector y.

Using these parameters, the combined Gaussian distribution of the clean vector x and the noisy vector y is expressed as shown in Equation (1). However, N (μ, Σ) represents a Gaussian distribution defined by the average μ and the covariance Σ.

  The Gaussian distribution calculation unit 133 will be described.

  The Gaussian distribution calculation unit 133 calculates a Gaussian parameter from the noise parameter and the prior distribution parameter by using unscented transformation, and outputs it to the Gaussian distribution storage unit 132.

Here, in calculating the Gaussian parameter, the nonlinear function y = f (x, n) for associating the clean vector x, the noise feature vector n, and the noisy vector y needs to be known. For example, when an MFCC vector is used as the feature vector, this nonlinear function is expressed as in Equation (2). However, the matrix C represents a discrete cosine transform, and its inverse matrix represents an inverse discrete cosine transform. Further, log and exp are assumed to act on each element of the vector.

  In the prior art disclosed in Non-Patent Document 1, Gaussian parameters are calculated using first-order Taylor approximation. On the other hand, in this embodiment, a Gaussian parameter is calculated using unscented transformation.

  In the following, the details of the prior art will be described first, and the problems will be pointed out. Then, the detail is demonstrated about the method of this embodiment.

  A conventional Gaussian parameter calculation method using first-order Taylor approximation will be described.

First, the nonlinear function of Expression (2) is approximated by first-order Taylor expansion as shown in Expression (3) below.

However, the matrices F and G are obtained by partial differentiation of the nonlinear function f by the clean vector x and the noise feature vector n, respectively, as shown in the following equation (4).

Further, the expansion point (x0, n0) of Taylor expansion is set to the prior average μx of the clean vector x and the average μn of the noise feature vector n, respectively, as shown in the following equation (5).

As described above, when the nonlinear function is approximated by the first order Taylor, the Gaussian parameter can be calculated by a linear operation. That is, the mean μy and covariance Σy of the noisy vector y, and the cross covariance Σxy of the clean vector x and the noisy vector y are calculated by the equations (6), (7), and (8), respectively.

  However, the above-described conventional method has a problem in that a Gaussian parameter calculation error is large due to an influence of an approximation error that occurs when a nonlinear function is approximated by a first-order Taylor.

  Next, a Gaussian parameter calculation method using unscented transformation according to the present embodiment will be described.

  “Unscented transformation” is a method of calculating a desired statistic with high accuracy in a nonlinear system. For details of unscented transformation, see Non-Patent Document 3 (S. Julier and J. Uhlmann, “Unscented filtering and nonlinear estimation,” Proceedings of the IEEE, vol. 92, no. 3, pp. 401-422, March. 2004.).

  This unscented conversion will be described.

  There is a first random variable x, and its mean μx and covariance Σx are known.

  There is a second random variable n, and its mean μn and covariance Σn are known.

  There is a third random variable y, and the third random variable y is calculated from the first random variable x and the second random variable n by a known nonlinear function y = f (x, n). And

  At this time, consider the problem of calculating the mean μy and covariance Σy of the third random variable y and the cross covariance Σxy between the first random variable x and the third random variable y. As a method for solving this problem with high accuracy, the above-mentioned unscented transformation is known.

  The Gaussian distribution calculation unit 133 calculates Gaussian parameters using this unscented transformation.

First, as shown in the following equation (9), a vector a obtained by connecting a clean vector x and a noise feature vector n is considered.

When the dimension of the clean vector x is Nx and the dimension of the noise feature vector n is Nn, the dimension of the vector a is Na = Nx + Nn. The average μa and the covariance Σa of the vector a are expressed by the following equations (10) and (11), respectively.

  Next, a set of samples called “sigma points” is generated. That is, p Na-dimensional vectors ai and weights wi associated with the Na-dimensional vectors ai are generated. Various methods are known as methods for generating sigma points. For example, Non-Patent Document 3 discloses them. Here, the “symmetric sigma point generation method” will be described. Any other sigma point generation method may be used.

In the symmetric sigma point generation method, p = 2Na vectors ai and weights wi associated therewith are generated as in the following Expression (12).

However, in Formula (12),

  Represents the i-th column (or row) of the square root of the matrix NaΣa.

  Next, the Gaussian distribution calculation unit 133 calculates yi for each of the p sigma points ai using the nonlinear function y = f (x, n). For example, when the feature vector is MFCC, the nonlinear function y = f (x, n) is expressed by Expression (2). Also, let xi be a vector obtained by extracting a portion corresponding to x of the i-th sample ai.

The gauss parameters to be calculated are calculated using xi and yi generated as described above (i = 1,... P). That is, the Gaussian distribution calculation unit 133 calculates the mean μy and the covariance Σy of the noisy vector y and the cross covariance Σxy of the clean vector x and the noisy vector y as shown in the following equations (14) to (16). To do.

  As described above, the Gaussian distribution calculation unit 133 calculates a Gaussian parameter from the prior distribution parameter and the noise parameter using the unscented transformation. In the conventional technique, the nonlinear function y = f (x, n) is approximated by the first-order Taylor expansion, so that the calculation error is large. However, the calculation error can be suppressed small by using unscented transformation.

  The calculation execution unit 134 will be described.

  The calculation execution unit 134 calculates a posterior distribution parameter from the noisy vector y based on the Gaussian parameter stored in the Gaussian distribution storage unit 132. As described above, the posterior distribution parameters are the posterior mean μx | y and the posterior covariance Σx | y.

When the two random variables x and y are distributed in a combined Gaussian distribution as shown in the equation (1), the clean vector x that is the first random variable when the noisy vector y that is the third random variable is observed. The following formula (17) for calculating the posterior average and the posterior covariance is known. The calculation execution unit 134 calculates the posterior distribution parameter using the equation (17).

  Next, the operation of the speech recognition apparatus 10 according to the present embodiment will be described with reference to FIG.

  First, in step S31, the feature extraction unit 11 calculates a noisy vector y from one frame of noisy speech.

  Next, in step S32, the noise estimation unit 12 estimates the noise parameter of the noise feature vector n from the noisy vector y.

  Next, in step S33, the Gaussian distribution calculation unit 133 calculates a Gaussian parameter using unscented transformation, and the Gaussian distribution storage unit 132 stores the Gaussian parameter.

  Next, in step S <b> 34, the calculation execution unit 134 calculates posterior distribution parameters based on the Gaussian parameters stored in the Gaussian distribution storage unit 132.

  Next, in step S35, the collation unit 14 collates the posterior distribution parameter of the clean vector x with a standard pattern of words stored in advance.

  Next, in step S36, the speech recognition apparatus 10 determines whether or not processing of all frames has been completed. If there are still frames that have not been processed, the process returns to step S31 to process the next frame. If all the frames have been processed, the process proceeds to step S37.

  Finally, in step S37, the collation unit 14 outputs the noisy speech word string based on the collation result.

  As described above, according to the present embodiment, since the Gaussian parameter can be calculated with high accuracy by using the unscented conversion, the feature enhancement effect can be enhanced and high speech recognition performance can be maintained even under noise.

(Second Embodiment)
Next, the speech recognition apparatus 10 according to the second embodiment will be described with reference to FIGS.

  In the speech recognition apparatus 10 according to the first embodiment, since the prior distribution of the clean vector x is expressed by a single Gaussian distribution, the prior distribution may not be expressed sufficiently precisely.

  Thus, in the speech recognition apparatus 10 of the present embodiment, the prior distribution of the clean vector x is expressed by a Gaussian mixture model, so that the prior distribution is expressed more precisely, so that feature enhancement works more effectively, and under noise Speech recognition performance is improved.

First, the Gaussian mixture model expressing the prior distribution of the clean vector x and its learning method will be described. First, in the present embodiment, M (M> 1) feature enhancement units 13 are provided. The prior distribution p (x) of the clean vector x is expressed by the following equation (18) using a Gaussian mixture model.

  Here, M (where M> 1) is the number of mixtures, k (where 1 <= k <= M) is the number of the feature enhancement unit 13, πk, μx (k), Σx (k) Represents the mixing weight, average, and covariance of the Gaussian distribution of the kth feature enhancement unit 13-k.

  In the first embodiment, the prior distribution is expressed by a single Gaussian distribution. However, in this embodiment, since the mixture of a plurality of Gaussian distributions is used, the prior distribution can be expressed more precisely.

  Gaussian mixture model parameters for representing the prior distribution of the clean vector x are previously learned from a corpus of clean speech and stored. Specifically, the Gaussian mixture model parameter of the above equation (18) is calculated by using an EM algorithm using a set of feature vectors extracted from a corpus of clean speech as learning data. Each feature enhancement unit 13 is generated so as to correspond to each phoneme, and calculates a Gaussian parameter corresponding to the phoneme for each feature enhancement unit 13.

  Next, the configuration of the speech recognition apparatus 10 of the present embodiment will be described with reference to FIG. FIG. 4 is a block diagram showing the voice recognition device 10.

  As shown in FIG. 4, the speech recognition apparatus 10 includes a feature extraction unit 11, a noise estimation unit 12, M feature enhancement units 13-1 to 13 -M, a weight calculation unit 41, an integration unit 42, a collation The unit 14 is provided. Since the feature extraction unit 11, the noise estimation unit 12, and the collation unit 14 are the same as those in the first embodiment, the same reference numerals are given and description thereof is omitted here.

  The feature enhancement unit 13 will be described.

  Each of the M feature emphasizing units 13-1,... 13-M is the same as the feature emphasizing unit 13 in the first embodiment. And different. The M feature emphasizing units 13-1,... 13-M have their own different parameters.

  That is, the prior distribution parameter storage unit 131-k included in the kth feature enhancement unit 13-k stores the kth Gaussian mixture model parameters μx (k) and Σx (k) of the Gaussian mixture model described above. .

  In addition, the Gaussian distribution calculating unit 133-k calculates the Gaussian parameters (μy (k), Σy (k), Σxy) from the noise parameters (μn and Σn) and the prior distribution parameters (μx (k) and Σx (k)). (K)) is calculated and stored in the Gaussian distribution storage unit 132-k.

  The calculation execution unit 134-k, based on the Gaussian parameter stored in the Gaussian distribution storage unit 132-k, the kth posterior distribution parameter, that is, the posterior mean μx | y (k) and the posterior covariance μx | y ( k) is calculated.

  The weight calculation unit 41 will be described.

  The weight calculation unit 41 calculates an integration weight used when integrating the outputs from the M feature emphasizing units 13-1,... 13-M. That is. Based on the Gaussian parameter calculated by each Gaussian distribution calculation unit 133-k, an integrated weight for each posterior distribution parameter is calculated for each frame.

Specifically, when the noisy vector y is observed, the posterior probability p (k | y) that the frame belongs to the feature enhancement unit 13-k is used as the integrated weight. The posterior probability p (k | y) is calculated by the following equation (19).

  πk is a mixing weight of the Gaussian mixture model described above. μy (k) and Σy (k) refer to the value of the Gaussian distribution storage unit 132-k of the k-th feature enhancement unit 13-k.

  The integration unit 42 will be described.

  The integrating unit 42 integrates outputs from the M feature emphasizing units 13-1 to 13 -M.

Specifically, the outputs μx | y (k) and Σx | y (k) from the M feature emphasizing units 13-1,... 13-M are integrated by the following equation (20), and μx | Y and Σx | y are output.

  Next, the operation of the speech recognition apparatus 10 according to the present embodiment will be described with reference to FIG. In addition, about the step same as 1st Embodiment, the same code | symbol is provided and description is performed simply.

  First, the feature extraction process in step S31 and the noise estimation process in step S32 are performed.

  Next, in step S33, the Gaussian distribution calculation unit 133-k of the feature enhancement unit 13-k calculates a Gaussian parameter using unscented transformation, and the Gaussian distribution storage unit 132-k stores the Gaussian parameter. To do.

  Next, in step S34, the calculation execution unit 134-k calculates a posterior distribution parameter based on the Gaussian parameter stored in the Gaussian distribution storage unit 132-k.

  Next, in step S51, the speech recognition apparatus 10 returns to step S33 if the processes for all the feature emphasizing units 13-1,... 13-M are not completed, and proceeds to step S52 if completed. Proceed with

  Next, in step S52, the weight calculation unit 41 calculates an integrated weight.

  Next, in step S53, the integration unit 42 integrates the outputs from the M feature extraction units 13-1, ... 13-M.

  Next, in step S35, the collation unit 14 collates with a standard pattern of words.

  Next, in step S36, the speech recognition apparatus 10 returns to step S31 if processing of all frames is not completed, and proceeds to step S37 if completed.

  Finally, in step S37, the collation unit 14 outputs the noisy speech word string based on the collation result.

  As described above, this embodiment can express the prior distribution more precisely by using the Gaussian mixture model than the case of using a single Gaussian distribution, enhances the feature enhancement effect, and has high speech even under noise. Recognition performance can be obtained.

(Third embodiment)
Next, the speech recognition apparatus 10 according to the third embodiment will be described with reference to FIGS.

  In the first and second embodiments, since the Gaussian parameter is calculated in all frames, the amount of calculation increases.

  Therefore, in this embodiment, in each frame, it is determined whether the Gaussian parameter needs to be calculated again. If it is determined that it is unnecessary, the calculation amount is reduced by omitting the recalculation of the Gaussian parameter. .

  Since the difference between the present embodiment and the first and second embodiments is only the configuration of the feature emphasizing unit 13, the description of other configurations is omitted.

  The feature emphasizing unit 13 of this embodiment will be described with reference to FIG. FIG. 6 is a block diagram of the feature emphasizing unit 13 in the present embodiment.

  The feature enhancement unit 13 includes a prior distribution parameter storage unit 131, a Gaussian distribution storage unit 132, a Gaussian distribution calculation unit 133, a calculation execution unit 134, a determination unit 61, and a first switch unit 62. Except for the determination unit 61 and the first switch unit 62, the second embodiment is the same as the first and second embodiments.

  The determination unit 61 will be described.

  The determination unit 61 determines whether it is necessary to recalculate Gaussian parameters in a certain frame.

  The noise parameter is input from the noise estimation unit 12 to the determination unit 61 for each frame. When the noise parameter changes greatly, the determination unit 61 determines that the Gaussian parameter needs to be recalculated because the value of the Gaussian parameter greatly varies accordingly. Conversely, when the noise parameter has not changed much, the value of the Gaussian parameter does not change much, so it is determined that recalculation of the Gaussian parameter is unnecessary.

  FIG. 7 is a block diagram of the determination unit 61. As illustrated in FIG. 7, the determination unit 61 includes a noise parameter storage unit 611, a fluctuation amount calculation unit 612, and a comparison unit 613.

  First, the noise parameter storage unit 611 stores the noise parameter of the frame for which the Gaussian distribution calculation unit 133 calculated Gaussian parameters at the end of the past.

The fluctuation amount calculation unit 612 calculates the fluctuation amount of the noise parameter from the current noise parameter input from the noise estimation unit 12 in the current frame and the past noise parameter stored in the noise parameter storage unit 611. For example, the fluctuation amount of the noise parameter is calculated by the Euclidean distance as shown in the following formula (21).

  Here, Δ is the fluctuation amount of the noise parameter, μn is the current noise parameter in the current frame, and μn with a superscript bar is a past noise parameter stored in the noise parameter storage unit 611.

  The comparison unit 613 compares the fluctuation amount with an arbitrary threshold value. If the fluctuation amount is larger than the threshold value, it is considered that the noise parameter has greatly fluctuated since the last calculation of the Gaussian parameter, and the Gaussian parameter needs to be recalculated. Output verdict.

  At the same time, a storage command is transmitted from the comparison unit 613 to the noise parameter storage unit 611, the current noise parameter in the current frame is stored in the noise parameter storage unit 611, and the past noise parameter is updated.

  If it is smaller than the threshold value, it is considered that the noise parameter has not fluctuated so much and a determination is made that recalculation of the Gaussian parameter is unnecessary. At this time, the value of the noise parameter storage unit 611 is not updated.

  The first switch unit 62 controls the operation of the Gaussian distribution calculation unit 133 according to the determination of the determination unit 61. That is, when it is determined that the Gaussian parameter needs to be recalculated, the Gaussian distribution calculating unit 133 is executed, the result is newly stored in the Gaussian distribution storage unit 132, and the calculation executing unit is used using the new Gaussian parameter. 134 calculates the posterior distribution parameters.

  On the other hand, when it is determined that recalculation is unnecessary, the first switch unit 62 omits the execution of the Gaussian distribution calculation unit 133. The contents of the Gaussian distribution storage unit 132 are not changed. The calculation execution unit 134 calculates the posterior distribution parameters using the past Gaussian parameters stored in the Gaussian distribution storage unit 132.

  In the case where a plurality of feature enhancement units 13-1,... 13-M as in the second embodiment are provided, each feature enhancement unit 13-1,. However, since the processing contents are the same, the single determination unit 61 can be shared by all the feature emphasizing units 13-1 to 13 -M.

  Next, the operation of the speech recognition apparatus 10 of this embodiment will be described with reference to FIG. FIG. 8 is a flowchart showing the operation of the speech recognition apparatus 10. Here, the operation of the configuration using the plurality of feature emphasizing units 13-1,... 13-M will be described.

  Since the operation of the configuration using the single feature emphasizing unit 13 as in the first embodiment is the same as that in a plurality of cases, the description thereof is omitted. Further, the same steps as those in the first or second embodiment will be given the same reference numerals and will be briefly described.

  First, the feature extraction process in step S31 and the noise estimation process in step S32 are performed.

  Next, in step S81, the determination unit 61 determines whether the Gaussian parameter needs to be recalculated based on the noise parameter fluctuation amount for the feature enhancement unit 13-k. If it is determined that it is necessary, a Gaussian parameter is calculated in step S33. If it is determined that it is unnecessary, the calculation of the Gaussian parameter is omitted.

  Next, in step S34, the calculation execution unit 134-k calculates a posterior distribution parameter based on the Gaussian parameter stored in the Gaussian distribution storage unit 132-k.

  Next, in step S51, the speech recognition apparatus 10 proceeds to step S52 if the processing has been completed for all the feature emphasizing units 13-1, ... 13-M. Otherwise, the process returns to step S81.

  Next, in step S52, the weight calculation unit 41 calculates an integrated weight.

  Next, in step S53, the integration unit 42 integrates the outputs from the M feature enhancement units 13-1, ... 13-M.

  Next, in step S35, the collation unit 14 collates with a standard pattern of words.

  Next, in step S36, the speech recognition apparatus 10 returns to step S31 if processing of all frames is not completed, and proceeds to step S37 if processing of all frames is completed.

  Finally, in step S37, the collation unit 14 outputs the noisy speech word string based on the collation result.

  As described above, in the present embodiment, it is determined whether or not the Gaussian parameter needs to be recalculated based on the fluctuation amount of the noise parameter, and the execution of the Gaussian distribution calculating unit 133 is omitted in the frame in which the recalculation is determined to be unnecessary. By doing so, the amount of calculation can be reduced.

(Fourth embodiment)
Next, the speech recognition apparatus 10 according to the fourth embodiment will be described with reference to FIGS.

  The present embodiment is intended to reduce the amount of calculation in the feature emphasizing unit 13 as in the third embodiment. That is, in the present embodiment, when the determination unit 61 determines that recalculation of the Gauss parameter is unnecessary, the simple calculation unit 91 that has a smaller calculation amount than the Gaussian distribution calculation unit 133 performs the calculation of the Gauss parameter. Update at least some parameters.

  Since the difference between the present embodiment and the third embodiment is only the configuration of the feature emphasizing unit 13, the description of other configurations is omitted.

  The feature emphasizing unit 13 of this embodiment will be described with reference to FIG. FIG. 9 is a block diagram of the feature enhancement unit 13.

  The feature enhancement unit 13 includes a prior distribution parameter storage unit 131, a Gaussian distribution storage unit 132, a Gaussian distribution calculation unit 133, a simple calculation unit 91, a determination unit 61, a second switch unit 92, and a calculation execution unit 134. Except for the simple calculation unit 91 and the second switch unit 92, the second embodiment is the same as the first to third embodiments.

  The simple calculation unit 91 updates at least a part of the Gaussian parameter with a smaller calculation amount than the Gaussian distribution calculation unit 133.

  Specifically, using the average μn that is one of the noise parameters (μn, Σn) in the current frame, the average μy value of the noise parameter that is one of the Gaussian parameters is expressed as μy = f (μx, μn). Calculated by The values of other Gaussian parameters (Σy, Σxy) are not calculated.

  Since the Gaussian distribution calculation unit 133 calculates an average μy and a Gaussian parameter (Σy, Σxy), which are one of noise parameters, using unscented transformation, there is a drawback in that the amount of calculation is large instead of being able to calculate the parameters with high accuracy. is there. On the other hand, the simple calculation unit 91 is less accurate but has a small amount of calculation. Therefore, the calculation amount in the feature enhancement unit 13 is suppressed by switching to the simple calculation unit 91 that can be executed with a small calculation amount for a frame that is determined to be unnecessary to calculate the Gaussian parameter based on the fluctuation amount of the noise parameter. Can do.

  Next, the operation of the speech recognition apparatus 10 of this embodiment will be described with reference to FIG. FIG. 10 is a flowchart showing the operation of the speech recognition apparatus 10.

  Here, the operation of the configuration using M feature emphasizing units 13-1,... 13-M will be described. Since the operation of the configuration using the single feature emphasizing unit 13 as in the first embodiment is the same as that in a plurality of cases, the description thereof is omitted. Further, the same steps as those in the first to third embodiments will be given the same reference numerals and will be briefly described.

  First, the feature extraction process in step S31 and the noise estimation process in step S32 are executed.

  Next, in step S81, the determination unit 61 determines whether the Gaussian parameter needs to be recalculated based on the noise parameter fluctuation amount for the feature enhancement unit 13-k. This determination is the same as in the third embodiment. If it is determined that recalculation is necessary, the operation of the Gaussian distribution calculation unit 133-k is executed in step S33. If it is determined that recalculation is unnecessary, the above-described operation of the simple calculation unit 91-k is executed in step S101.

  Next, in step S34, the calculation execution unit 134-k calculates a posterior distribution parameter based on the Gaussian parameter stored in the Gaussian distribution storage unit 132-k.

  Next, in step S51, the speech recognition apparatus 10 proceeds to step S52 if the processing has been completed for all the feature emphasizing units 13-1, ... 13-M. If not completed, the process returns to step S81.

  Next, in step S52, the weight calculation unit 41 calculates an integrated weight.

  Next, in step S53, the outputs from the M feature emphasizing units 13-1, ... 13-M are integrated.

  Next, in step S35, the collation unit 14 collates with a standard pattern of words.

  Next, in step S36, the speech recognition apparatus 10 returns to step S31 if processing of all frames is not completed, and proceeds to step S37 if processing of all frames is completed.

  Finally, in step S37, the collation unit 14 outputs the noisy speech word string based on the collation result.

  As described above, in the present embodiment, it is determined whether the recalculation of the Gaussian parameter is necessary or unnecessary based on the fluctuation amount of the noise parameter, and the simple calculation unit 91 having a small calculation amount is determined in the frame in which the recalculation is determined to be unnecessary. The amount of calculation can be reduced by switching to.

(Example of change)
The present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the gist thereof.

It is a block diagram which shows the structure of the speech recognition apparatus of 1st Embodiment. It is a block diagram which shows the structure of the characteristic emphasis part of 1st Embodiment. It is a flowchart which shows operation | movement of the speech recognition apparatus of 1st Embodiment. It is a block diagram which shows the structure of the speech recognition apparatus of 2nd Embodiment. It is a flowchart which shows operation | movement of the speech recognition apparatus of 2nd Embodiment. It is a block diagram which shows the structure of the characteristic emphasis part of 3rd Embodiment. It is a block diagram which shows the structure of a determination part. It is a flowchart which shows operation | movement of the speech recognition apparatus of 3rd Embodiment. It is a block diagram which shows the structure of the characteristic emphasis part of 4th Embodiment. It is a flowchart which shows operation | movement of the speech recognition apparatus of 4th Embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Speech recognition apparatus 11 Feature extraction part 12 Noise estimation part 13 Feature emphasis part 14 Collation part 41 Weight calculation part 42 Integration part 61 Determination part 62 1st switch part 91 Simple calculation part 92 2nd switch part 131 Prior distribution parameter memory | storage part 132 Gaussian distribution storage unit 133 Gaussian distribution calculation unit 134 Calculation execution unit 611 Noise parameter storage unit 612 Variation amount calculation unit 613 Comparison unit

Claims (6)

A feature extraction unit that extracts a noisy speech feature vector for each frame from the input noisy speech;
A noise estimator for estimating a noise feature distribution parameter of a noise feature vector related to noise superimposed on the noisy speech;
A prior distribution parameter storage unit for storing a prior distribution parameter of a clean speech feature vector related to clean speech;
A Gaussian distribution calculation unit that calculates a combined Gaussian distribution parameter of the clean speech feature vector and the noisy speech feature vector for each frame using an unscented transform from the noise feature distribution parameter and the prior distribution parameter;
A calculation execution unit that calculates a posterior distribution parameter of the clean speech feature vector for each frame from the noisy speech feature vector using the combined Gaussian distribution parameter;
Collating the posterior distribution parameter with a pre-stored word standard pattern for each frame, and outputting the noisy speech word string based on the collation result;
A speech recognition apparatus comprising: A plurality of the prior distribution parameter storage unit, the Gaussian distribution calculation unit, and the calculation execution unit, respectively,
Based on the combined Gaussian distribution parameters respectively calculated by the respective Gaussian distribution calculating units, a weight calculating unit that calculates an integrated weight for each posterior distribution parameter for each frame;
Integrating each posterior distribution parameter based on each integration weight, and outputting the integrated posterior distribution parameter to the matching unit for each frame;
The speech recognition apparatus according to claim 1, further comprising: A Gaussian distribution storage unit that stores the combined Gaussian distribution parameter calculated by the Gaussian distribution calculation unit for each frame;
The fluctuation amount of the noise feature distribution parameter is obtained for each frame, and when the fluctuation amount is smaller than an arbitrary threshold, it is determined that recalculation of the combined Gaussian distribution parameter is unnecessary, and when the fluctuation amount is larger than the threshold. Is a determination unit that determines that recalculation of the combined Gaussian distribution parameter is necessary;
(1) For the frame determined to require recalculation, the combined Gaussian distribution parameter recalculated by the Gaussian distribution calculation unit is sent to the calculation execution unit, and (2) it is determined that recalculation is unnecessary. For the frame, a first switch unit that sends the combined Gaussian distribution parameter calculated for the previous frame before the frame calculated by the Gaussian distribution calculation unit to the calculation execution unit;
The speech recognition apparatus according to claim 1, further comprising: A Gaussian distribution storage unit that stores the combined Gaussian distribution parameter calculated by the Gaussian distribution calculation unit for each frame;
The fluctuation amount of the noise feature distribution parameter is obtained for each frame, and when the fluctuation amount is smaller than an arbitrary threshold, it is determined that recalculation of the combined Gaussian distribution parameter is unnecessary, and when the fluctuation amount is larger than the threshold. Is a determination unit that determines that recalculation of the combined Gaussian distribution parameter is necessary;
From the noise feature distribution parameter and the prior distribution parameter, a simple calculation unit that calculates one parameter among the combined Gaussian distribution parameters for each frame,
(1) For the frame determined to require recalculation, the combined Gaussian distribution parameter recalculated by the Gaussian distribution calculation unit is sent to the calculation execution unit, and (2) it is determined that recalculation is unnecessary. For the frame, a second switch unit that sends the combined Gaussian distribution parameter excluding the one parameter calculated in the simple calculation unit and the one parameter stored in the Gaussian distribution storage unit to the calculation execution unit;
The speech recognition apparatus according to claim 1, further comprising: A feature extraction step for extracting a noisy speech feature vector for each frame from the input noisy speech;
A noise estimation step of estimating a noise feature distribution parameter of a noise feature vector related to noise superimposed on the noisy speech;
Using the noise feature distribution parameter and the pre-distribution parameter of the clean speech feature vector relating to the clean speech stored in advance, the combined Gaussian distribution parameter of the clean speech feature vector and the noisy speech feature vector is converted into the frame using unscented transformation. A Gaussian distribution calculating step to calculate for each,
A calculation execution step of calculating a posterior distribution parameter of the clean speech feature vector for each frame from the noisy speech feature vector using the combined Gaussian distribution parameter;
Collating the posterior distribution parameter with a pre-stored word standard pattern for each frame, and outputting the noisy speech word string based on the collation result;
A speech recognition method comprising: A feature extraction function that extracts a noisy speech feature vector from the input noisy speech for each frame;
A noise estimation function for estimating a noise feature distribution parameter of a noise feature vector related to noise superimposed on the noisy speech;
Using the noise feature distribution parameter and the pre-distribution parameter of the clean speech feature vector relating to the clean speech stored in advance, the combined Gaussian distribution parameter of the clean speech feature vector and the noisy speech feature vector is converted into the frame using unscented transformation. Gaussian distribution calculation function to calculate every,
A calculation execution function for calculating the posterior distribution parameter of the clean speech feature vector for each frame from the noisy speech feature vector using the combined Gaussian distribution parameter;
A collation function for collating the posterior distribution parameter with a pre-stored word standard pattern for each frame and outputting the word sequence of the noisy speech based on the collation result;
Is a speech recognition program that realizes a computer.